Complexity Bias

Transcript

1. Molly Lewis Stanford University 15 December 2015 The role of communicative pressures in shaping the lexicon
2. However, limits to arbitrariness (Köhler, 1929; Maurer, et al., 2006; Ramachandran & Hubbard, 2001; Farmer, Christiansen, & Monaghan, 2006; Zipf, 1936; Piantadosi, Tily, & Gibson, 2011) horse “The linguistic sign is arbitrary” – Saussure (1916) !kalë !ناصح !ձի !at !zaldi !конь ! ঘা#া !konj !кон !cavall!kabayo !
   !
   !konj !kůň !hest !paard!ĉevalo !hobune !kabayo !hevonen !cheval !cabalo !Pferd !άλογο!ઘોડો !chwal!doki !סוס !घोड़ा !nees !ló !hestur !anyịnya !kuda !capall!cavallo !
   !jaran !!"# !សេះ !݈ມ"າ !equo !zirgs !arklys!коњ !kuda !! !hoiho!घोडा !адуу !घोडा !hest !بسا !koń !cavalo !ਘ"ੜਾ !cal !лошадь !коњ !kôň !konj !faras !caballo !farasi!häst !!"ை !!ర#$ !ม้า !кінь !اڑوھگ !ngựa !ceffyl !!
3. Complexity Bias A bias to map longer words (in terms of phonemes, morphemes, syllables) to more complex referents tupabugorn
4. Complexity Bias Theories of communication predict tradeoff between length and predictability Horn Implicatures (Horn, 1984) I  turned  on  the  car. I  got  the  car  to  turn  on. TYPICAL ATYPICAL Uniform Information Density (Aylett & Turk, 2004; A. Frank & Jaeger, 2008)
5. Outline I. Do participants have a productive complexity bias? – Novel real objects (Study 1) – Artificial objects (Study 2) II. What is complexity? (Study 3) III. Is there a complexity bias in the lexicon? – English (Study 4) – Cross-linguistically (Study 5) IV. Where does the lexical bias come from? (Study 6)
6. Study 1a: Explicit complexity judgment
7. N=60 Least complex Most complex
8. Study 1b: Mapping task Map novel word to novel object, given 2 alternatives
9. Study 1b: Design Referent complexity x word length (within subject) Linguistic stimuli: – short words (e.g., "bugorn,” "ratum,” "lopus”) – long words (e.g., "tupabugorn,” "gaburatum,” "fepolopus") Referent stimuli: – Divided objects into quintiles, based on explicit complexity norms – Tested every pairing of quintiles (15 conditions): 1/1, 1/2, 1/3, 1/4, 1/5, 2/2, 2/3, etc. Procedure: 8 trials/participant
10. Example 1/5 Trial Quintile 1 2 3 4 5
11. Study 1b: Results • • • • • • • • • • • • • • • 1/1 1/2 1/3 1/4 1/5 2/2 2/3 2/4 2/5 3/3 3/4 3/5 4/4 4/5 5/5 r= −0.7 −0.25 0.00 0.25 0.50 0.50 0.75 1.00 complexity rating ratio effect size (cohen's d) N = 1500 Target biased to have long label Target less complex
12. Evidence for a productive complexity bias in an online mapping task But: manipulate complexity correlationally (difficult to interpret causation) Study 2: Direct manipulation of complexity Is there a productive complexity bias?
13. Quintile 1 2 3 4 5 Example 1/5 Trial Study 2: Stimuli
14. Study 2: Results N = 750 Target biased to have long label Target less complex • • • • • • • • • • • • • • • 1/1 1/2 1/3 1/4 1/5 2/2 2/3 2/4 2/5 3/3 3/4 3/5 4/4 4/5 5/5 r= −0.87 −0.25 0.00 0.25 0.50 0.25 0.50 0.75 1.00 1.25 complexity rating ratio effect size (cohen's d)
15. Evidence for a productive complexity bias in online mapping task – Manipulating complexity both correlationally and directly Complexity quantified in terms of visual complexity But: What is the underlying complexity construct? What is complexity?
16. In visual cognition, use processing time as index of information load (Alvarez & Cavanaugh, 2004) – more information requires more processing time – not perfect measure, but expect monotonic relationship – search rate task < What is complexity?
17. Recognition memory task measure study time per object (30 objects) (60 objects) Study 3: Implicit complexity judgment
18. • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • r= 0.52 7.0 7.2 7.4 7.6 0.00 0.25 0.50 0.75 1.00 Object Complexity Norms Log RT (ms) Novel object complexity norms NRT =  494 NC =  60
19. Study 3a: Novel Real Objects • • • • • • • • • • • • • • • 1/1 1/2 1/3 1/4 1/5 2/2 2/3 2/4 2/5 3/3 3/4 3/5 4/4 4/5 5/5 r= −0.7 −0.25 0.00 0.25 0.50 0.50 0.75 1.00 complexity rating ratio effect size (cohen's d) Complexity Norms • • • • • • • • • • • • • • • 1/1 1/2 1/3 1/4 1/5 2/2 2/3 2/4 2/5 3/3 3/4 3/5 4/4 4/5 5/5 r= −0.71 −0.25 0.00 0.25 0.50 0.985 0.990 0.995 1.000 RT ratio effect size (cohen's d) RT Norms
20. • • • • • • • • • • • • • • • 1/1 1/2 1/3 1/4 1/5 2/2 2/3 2/4 2/5 3/3 3/4 3/5 4/4 4/5 5/5 r= −0.8 −0.25 0.00 0.25 0.50 0.95 0.96 0.97 0.98 0.99 1.00 RT ratio effect size (cohen's d) Study 3b: Artificial Objects • • • • • • • • • • • • • • • 1/1 1/2 1/3 1/4 1/5 2/2 2/3 2/4 2/5 3/3 3/4 3/5 4/4 4/5 5/5 r= −0.87 −0.25 0.00 0.25 0.50 0.25 0.50 0.75 1.00 1.25 complexity rating ratio effect size (cohen's d) Complexity Norms RT Norms
21. Exp. 1-2: suggest a productive complexity bias with novel words Exp. 3: Complexity bias related to processing time. Next: Is this bias present in natural languages? Study 4: Explicit complexity norms for English words Is this bias in natural language?
22. Study 4: English complexity norms Rate words for complexity alphabet
23. Complexity norms Normed 499 English words 30 words/participant N = 250 participants Word Lengths Word Length (characters) Frequency 2 4 6 8 10 12 0 20 40 60 80 100 120 140
24. Study 4: Results r CL ŸF = .60 N  =  250 Characters: Phonemes: r CL = .69 r CL ŸF = .61 Syllables: r CL = .67 r CL ŸF = .58 r= 0.69 1 2 3 4 5 6 7 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Word Length (characters) Complexity Rating Reliable controlling for concreteness, familiarity and imagability
25. Study 5: Cross-linguistic Evidence that complexity is related to length in English (controlling for other semantic variables) But: does this extend to other languages? Examined relationship between word lengths for normed words in 80 languages Google translate – Native speakers hand-checked 12 languages – Accuracy: 92%
26. 0.0 0.2 0.4 0.6 english afrikaans maltese danish norwegian macedonian yiddish dutch russian serbian croatian portuguese espernto galician basque bosnian welsh armanian italian swedish georgian belarusian icelandic estonian bulgarian german hungarian latvian ukranian spanish thai french nepali polish chinese czech hmong slovenian slovak mongolian hindi zulu vietnamese finnish swahili irish lao hausa filipino lithuanian haitian.creole romanian khmer punjabi catalan gujarati indonesian greek hebrew azerbaijani malay cebuana javanese albanian kanada turkish yoruba maori somali korean telugu urdu tamil bengali arabic latin japanese igbo persian marathi Language Pearson's r 0.0 0.2 0.4 0.6 english afrikaans maltese danish norwegian macedonian yiddish dutch russian serbian croatian portuguese espernto galician basque bosnian welsh armanian italian swedish georgian belarusian icelandic estonian bulgarian german hungarian latvian ukranian spanish thai french nepali polish chinese czech hmong slovenian slovak mongolian hindi zulu vietnamese finnish swahili irish lao hausa filipino lithuanian haitian.creole romanian khmer punjabi catalan gujarati indonesian greek hebrew azerbaijani malay cebuana javanese albanian kanada turkish yoruba maori somali korean telugu urdu tamil bengali arabic latin japanese igbo persian marathi Language Pearson's r 0.0 0.2 0.4 0.6 english afrikaans maltese danish norwegian macedonian yiddish dutch russian serbian croatian portuguese espernto galician basque bosnian welsh armanian italian swedish georgian belarusian icelandic estonian bulgarian german hungarian latvian ukranian spanish thai french nepali polish chinese czech hmong slovenian slovak mongolian hindi zulu vietnamese finnish swahili irish lao hausa filipino lithuanian haitian.creole romanian khmer punjabi catalan gujarati indonesian greek hebrew azerbaijani malay cebuana javanese albanian kanada turkish yoruba maori somali korean telugu urdu tamil bengali arabic latin japanese igbo persian marathi Language Pearson's r # open class words = 453 Correlation between complexity norm and word length
27. • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • Cross−Linguistic Complexity Bias Geographical distribution of complexity bias
28. Complexity bias by language family 0.0 0.1 0.2 0.3 0.4 Basque Kartvelian Indo−European Uralic Sino−Tibetan Tai−Kadai Hmong−Mien Austro−Asiatic Creoles and Pidgins Afro−Asiatic Altaic Austronesian Niger−Congo Korean Dravidian Japanese Language Family Pearson's r Complexity bias by language family
29. 0.0 0.2 0.4 0.6 english afrikaans maltese danish norwegian macedonian yiddish dutch russian serbian croatian portuguese espernto galician basque bosnian welsh armanian italian swedish georgian belarusian icelandic estonian bulgarian german hungarian latvian ukranian spanish thai french nepali polish chinese czech hmong slovenian slovak mongolian hindi zulu vietnamese finnish swahili irish lao hausa filipino lithuanian haitian.creole romanian khmer punjabi catalan gujarati indonesian greek hebrew azerbaijani malay cebuana javanese albanian kanada turkish yoruba maori somali korean telugu urdu tamil bengali arabic latin japanese igbo persian marathi Language Pearson's r Where does the bias in language come from? Complexity bias in individual speakers over time leads to the same regularity emerging in the structure of the lexicon Productive complexity bias Lexical complexity bias
30. Where does the bias in language come from? A mechanism over multiple timescales (Griffiths & Kalish, 2007) Conversational timescale (minutes) Language change timescale (many years) t
31. Word learning Study 6: Lexical learning (10 word-object pairs x4) Word recall
32. Study 6a: Lexical learning Generated random lexicon – Words: 3, 5, 7, 9, 11 characters (CV syllables) – Objects: 2 from each complexity quintile Predictions: - Word forms: Become more stable - Complexity bias: Shorten words for simple objects, lengthen words for complex objects ninop nin ninop ninopen
33. Study 6a: Results Words are shortened for simple objects • • • • • −0.5 0.0 0.5 1.0 1.5 1 2 3 4 5 Complexity quintile Number characters removed Shorter words Complexity N = 50
34. Study 6b: Iterated lexical learning Iterated learning paradigm – a method for simulating language change Gave the labels generated by participants to a new set of participants Iterated for total of 7 generations 50 participants/generation
35. Study 6b: Iterated lexical learning Word forms become more stable Across generations, words tended to: ①become easier to remember ②shorten ③increase in bigram transitional probability ④decrease in variability ⑤decrease in word change (Levenshtein edit distance) • • • • • • • • • • • • • • 0.00 0.25 0.50 0.75 1.00 1 2 3 4 5 6 7 Generation Proportion correct Mean accuracy
36. Study 6b: Iterated lexical learning Complexity bias persists across time 1 2 3 4 5 6 7 • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • 0 1 2 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 Complexity quintile Cumulative characters removed Complexity bias across generations Shorter words complexity
37. But, why doesn’t it strengthen? Pressure to simplify suppresses complexity bias Two competing communicative pressures (Horn, 1984): – Speaker/learner pressure à compression – Listener pressure à differentiation Task not communicative – Presence of a listener pressure reduces compression? – Currently running version with interlocutor, and version with binary feedback
38. Conclusion Evidence for: – a complexity bias in the lexicon – productive – related to a basic cognitive process – can emerge from learning biases Suggests: – complexity as constraint on arbitrariness in language – cognitive biases are reflected in the structure of the lexicon – communicative biases may shape the lexicon
39. Thank you
40. The learnability pressure at at the language- change timescale Learnability pressure as one factor influencing the morphological complexity of a language (Lupyan & Dale, 2010) Languages adapt to their particular social context à simpler if acquired by diverse population Predicts: Languages with more speakers should have smaller complexity bias.
41. Evidence for a relationship between learnability pressure and complexity bias • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • r=−0.34 0.1 0.2 0.3 0.4 0.5 12.5 15.0 17.5 20.0 Log Population (million) Complexity Bias (Pearson's r) Language Family • • • • • • • • • • • • • • • • Afro−Asiatic Altaic Austro−Asiatic Austronesian Basque Creoles and Pidgins Dravidian Hmong−Mien Indo−European Japanese Kartvelian Korean Niger−Congo Sino−Tibetan Tai−Kadai Uralic
42. 0.00 0.25 0.50 0.75 1.00 1 2 3 4 5 Number of syllables Proportion selection complex object
43. 0.00 0.25 0.50 0.75 1.00 1 2 3 4 5 Number of syllables Proportion selection complex object glm(responseValue ~  len,  data=dc,  family  =  "binomial”)